How the work is actually done — and how we catch our own mistakes
Every result on this platform comes from a specific, repeatable procedure, and every number is labeled with where it came from. This page lays out those procedures in plain language: how HER2T reads and reasons, how it computes on real protein structure, how the GPU molecular-dynamics runs work, and — most importantly — the discipline that lets the system be graded honestly, including a worked example of a reproducibility failure we detected and corrected rather than published.
Evidence tiers — always labeled, never mixed
Inferred · AI reasoning
Generated by HER2T reasoning from known structural facts. A hypothesis, not a physics calculation and not a measurement.
Computed · deterministic calculation
A deterministic calculation performed directly on real atomic coordinates (interface contacts, distances, buried surface area, clashes). Reproducible math on structure — not AI reasoning, not a stochastic simulation, and not a measurement.
Simulated · trajectory / stochastic physics
Produced by an actual trajectory or stochastic physics calculation (molecular dynamics, docking, free-energy perturbation). Reserved — no result in HER2T currently earns this tier.
Measured · experimental data
A real measurement from the physical world, backed by a citation. The only tier that reflects reality rather than a model of it.
The HER2T Compute Lab is intentionally paused under a safety freeze. No simulations are being sent to any GPU node.
No molecular-docking engine is attached. Docking stages stay blocked, never faked.
No FEP / binding-free-energy engine is attached. Those stages are not run.
Physical experiments go to qualified partner laboratories today; a dedicated HER2T lab is the goal. No physical experiment has been run yet.
1 · Reading and reasoning (inferred tier)
HER2T continuously reads the HER2/ERBB2 literature and turns it into structured, checkable objects. This stage is AI reasoning over published science — powerful, but explicitly the inferred tier. It is a source of hypotheses, never of proof.
Extract claims
Each paper is read and its factual assertions are pulled out as individual, attributable claims tied back to their source.
Find contradictions
Claims are compared across the whole corpus to surface where the published science genuinely disagrees with itself.
Raise sharper questions
Every unresolved contradiction is turned into one or more tracked, answerable questions.
Form and stress-test hypotheses
Candidate explanations are generated, reviewed by a skeptic pass, and ranked — the weak ones are discarded before they ever reach a prediction.
Freeze a testable prediction
The strongest ideas become predictions with a measurable consequence, frozen into the permanent ledger before any evidence is attached (see section 5).
2 · Deterministic structural computation (computed tier)
The computed tier is reproducible arithmetic performed directly on real Protein Data Bank atomic coordinates — no language model produces these numbers. Given the same structure and the same parameters, anyone re-running the calculation gets the same answer.
The in-house engine (HER2T-Phys) measures things like interface contacts between HER2 and a bound antibody, distances around a mutation site, buried surface area, and steric clashes. The raw numbers are stored verbatim alongside the exact parameters used (atom-selection, distance cutoffs, probe radii, chains), so a result can always be reproduced or audited. Any language-model reading of those numbers is stored as a clearly separate interpretation — the numbers themselves are never model-generated.
3 · GPU molecular dynamics — the Compute Lab
The Compute Lab is currently paused under a safety freeze, and no simulations are running. When it runs, molecular-dynamics simulations run on a GPU node (an NVIDIA RTX 4090) using OpenMM. This is genuine, physically-grounded computation — but it is still computation, not a wet-lab measurement, and we say so everywhere it appears.
The procedure for each simulated system
- • Solvate the protein in explicit TIP3P water with a 1.0 nm padding box.
- • Treat long-range electrostatics with Particle-Mesh Ewald (PME).
- • Integrate with a Langevin-middle thermostat at 310 K and a 2 fs timestep.
- • Energy-minimize, equilibrate (NVT, 100 ps), then run 10 ns of production dynamics.
- • For a mutation study, run the wild-type and the mutant through the identical protocol and compare.
Typical wall-clock time is roughly two to two-and-a-half hours per mutation study on the 4090. You can watch jobs move through this pipeline on the reproduction ladder and inspect completed runs in the Compute Lab.
4 · What is deliberately not connected yet
Honesty cuts both ways — the platform is explicit about what it cannot do. Two things are intentionally not wired up, and any step that needs them stops and says so rather than faking a result:
Molecular docking
The Compute Lab runs molecular dynamics, not docking. Docking is the next backend to attach; until it is, reproduction stages that require it are shown as explicitly blocked.
The Experimental (wet) Lab
No prediction here has been tested at the bench yet. Physical experiments go to qualified partner laboratories today, with a dedicated HER2T lab as the goal. The number of wet-lab-confirmed predictions is currently zero, and the ledger shows it.
5 · The reproducibility discipline
These rules are what make the platform gradeable instead of just persuasive:
Predictions are frozen before evidence. A prediction’s wording, direction, confidence, and timestamp become immutable the moment it enters the Prediction Ledger. The system can never quietly rewrite a claim after seeing how it turned out.
The ledger is append-only. Misses are kept and counted on the accountability scoreboard, not deleted. An honest score — including the wrong predictions — is the entire point.
Evidence tiers are never mixed. A reasoned hypothesis, a deterministic calculation, a simulation, and a laboratory measurement are labeled distinctly everywhere. Computational support is never dressed up as biological validation.
Results are sanity-checked, and failures are surfaced. When a computed number is physically implausible, the pipeline flags it, refuses to draw a scientific conclusion from it, and — where possible — diagnoses and re-runs. The next section is a real example.
Worked example: the 28.7 Å RMSD that was a bug, not a discovery
One of the first real molecular-dynamics runs — a study of the L755S kinase-domain mutation against wild-type HER2 — produced a wild-type Cα RMSD of about 28.7 Å over 10 ns. Taken at face value, that would suggest the protein had massively unfolded.
What a careless pipeline would have done
Reported “the mutation destabilizes HER2” — a dramatic, publishable, and completely wrong biological conclusion.
What actually happened
A stable folded domain does not drift 28.7 Å in 10 ns. That value tripped the built-in sanity threshold, so the pipeline refused to treat it as a real structural change and flagged the run for diagnosis.
The real cause — a periodic-boundary imaging artifact
Simulations run in a repeating box. When the protein diffuses across a box edge, a naive RMSD measures it as if it were stretched across the whole box — an artifact of how the coordinates were imaged, not any real motion of the molecule. We confirmed the mechanism with a controlled test: a deliberately stable structure (true RMSD ≈ 0 Å) reproduced the same ~24–28 Å reading once its coordinates were wrapped across the boundary.
The fix
The trajectory writer now records whole molecules, and the analysis step reconstructs whole molecules and slices to the protein before computing RMSD. Because the GPU node’s scratch disk is wiped after each job, the original trajectory no longer existed to re-analyze, so the honest fix was to re-run the study on the corrected pipeline rather than quietly patch a stored number.
The verdict on the original run
Recorded as inconclusive — and never overwritten. The flawed number stays in the record as what it was; the corrected re-run is a separate, clearly-labeled result.
The point of this example is not that a bug occurred — bugs always will. The point is that the system was built to distrust its own output, catch the implausible number, explain it, and correct it in the open, instead of turning an artifact into a headline.